|Publication number||US7161906 B2|
|Application number||US 10/017,173|
|Publication date||Jan 9, 2007|
|Filing date||Dec 14, 2001|
|Priority date||Dec 15, 2000|
|Also published as||US20020075883|
|Publication number||017173, 10017173, US 7161906 B2, US 7161906B2, US-B2-7161906, US7161906 B2, US7161906B2|
|Inventors||Martin S. Dell, Zbigniew M. Dziong, Wei Li, Yu-Kuen Ouyang, Matthew Tota, Yung-Terng Wang|
|Original Assignee||Agere Systems Inc.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (34), Referenced by (60), Classifications (24), Legal Events (5)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application claims the benefit of the filing date of U.S. provisional application No. 60/255,749, filed on Dec. 15, 2000. This application is one of a set of U.S. patent applications consisting of Ser. Nos. 10/017,173, 10/017,174, and 10/017,503, all of which were filed on the same date and the teachings of all of which are incorporated herein by reference.
1. Field of the Invention
The present invention relates to data communications, and, in particular, to switch fabrics for routing packets of user data between multiple sources and destinations.
2. Description of the Related Art
In non-blocking switch fabrics, user data received from any particular source 102 can be routed through switch fabric 100 to any particular destination 110, independent of whether any other user data is being routed from one or more other sources towards one or more other destinations through switch fabric 100. When packets of user data transmitted from the sources are received at an input port of input stage 104, the user data is buffered for eventual transmission through switching stage 106 to output stage 108, where the routed user data is again buffered for eventual transmission to the destination via an output port of the output stage associated with that destination.
To perform these data buffering functions, input stage 104 and output stage 108 are both configured with a number of different routing queues (e.g., designated memory locations) in which the user data are temporarily stored. In prior art switch fabrics, such as switch fabric 100 of
The input stage is implemented with a scheduler that is responsible for determining when to transmit data from the various queues in the input stage to the switching stage for routing. When the scheduler determines that a particular queue is to be serviced (i.e., have user data transmitted to the switching stage), a bid (i.e., service request) is transmitted from the input stage (e.g., encoded in the header of a cell carrying other user data in its payload) to a bid arbitrator implemented as part of the switching stage. The bid arbitrator determines whether to accept (i.e., grant) or reject (i.e., deny) the bid received from the input stage. A message is transmitted from the switching stage back to the input stage identifying whether the bid was accepted or rejected. If a bid is accepted, then a connection is eventually established within the switching stage for routing the user data associated with that granted bid. In response to receiving a granted bid message, user data is transmitted from the queue in the input stage to the switching stage for routing through the corresponding connection to the appropriate queue in the output stage.
In prior art switch fabrics, such as switch fabric 100 of
Depending on the implementation, a bid arbitrator implemented in the switching stage may use different types of information in determining whether to accept or reject a particular bid. For example, a bid arbitrator may rely on “fairness” when considering a particular bid, where “fairness” relates to the notion of equitably sharing over time the resources of the switching stage among all of the different queues in the input stage that desire service. In addition, a bid arbitrator of the prior art might be designed to take into account the status of the various queues in the output stage. In particular, if a particular queue in the output stage is currently full, then a bid arbitrator may use that information (which is transmitted from the output stage to the switching stage) to reject a current bid for service that would result in even more user data being transmitted to that queue.
With the explosive demand for bandwidth and the rapid advances in electronics and devices, communication networks are going through rapid expansion and stressing the scalability of switches and routers. In the foreseeable future, packet services using electronics switching are cost competitive for switch capacity up to tens of terabits per second (Tbps). Depending on the level of aggregations and the traffic, it is also desirable to have a switch fabric that can scale from gigabits per second (Gbps) to Tbps and can support effectively a widely ranging number, say tens to thousands, of ports. Further, the ability to support (a) quality of service (QoS) for a variety of services and applications and (b) various network protocols such as IP, ATM, and MPLS is highly desirable.
To that end, embodiments of the present invention are directed to a scalable multi-Tbps protocol-independent switch fabric architecture that meets such a challenge. The architecture is cost effective for building switches supporting wire-speed switching up to 1024 ports with capacity ranging from tens of Gbps to tens of Tbps and different QoS traffic classes. Scalability is achieved via self-routing of cells with a routing queue number scheme coupled with an efficient back-pressure and flow control. Arbitrarily high utilization of the crossbars can be achieved using an efficient output buffering and re-sequencing scheme.
In one embodiment, the present invention is a switch fabric for routing data from one or more sources towards one or more destinations, comprising a switching stage configured between an input stage and an output stage, wherein the input stage is configured to receive the data transmitted from the one or more sources and forward the data to the switching stage; the switching stage is configured to route the data received from the input stage to the output stage; the output stage is configured to transmit the data received from the switching stage towards the one or more destinations; the input stage has one or more input ports, wherein each input port can be configured to receive data from a source; the output stage has one or more output ports, wherein each output port can be configured to transmit data towards a destination; the input stage comprises a plurality of input routing queues configured to store the data until the data is ready to be forwarded to the switching stage; the output stage comprises a plurality of output routing queues configured to store the data until the data is ready to be transmitted towards the one or more destinations; at least one input port can be programmably configured to store data in two or more input routing queues that are associated with a single output port; and at least one output port can be programmably configured to receive data from two or more output routing queues that are associated with a single input port.
In another embodiment, the present invention is a switch fabric for routing data from one or more sources towards one or more destinations, comprising a switching stage configured between an input stage and an output stage wherein the input stage is configured to receive the data transmitted from the one or more sources and forward the data to the switching stage; the switching stage is configured to route the data received from the input stage to the output stage; the output stage is configured to transmit the data received from the switching stage towards the one or more destinations; the input stage transmits bids to the switching stage to request connections through the switching stage for routing the data; the output stage is configured to transmit status information about the output stage to the input stage; and the input stage is configured to generate the bids transmitted to the switching fabric based on the status information about the output stage.
In yet another embodiment, the present invention is a switch fabric for routing data from one or more sources towards one or more destinations, comprising a switching stage configured between an input stage and an output stage wherein the input stage is configured to receive the data transmitted from the one or more sources and forward the data to the switching stage; the switching stage is configured to route the data received from the input stage to the output stage; the output stage is configured to transmit the data received from the switching stage towards the one or more destinations; the input stage is configured to transmit bids to the switching stage to request connections through the switching stage for routing the data to the output stage; the switching stage is configured to determine whether to accept or reject each bid and to transmit a grant/rejection signal to the input stage identifying whether the bid is accepted or rejected; for each rejected bid, the grant/rejection signal explicitly identifies a reason for rejecting the bid; and the input stage determines how to react to a rejected bid based on the reason the bid was rejected.
Other aspects, features, and advantages of the present invention will become more fully apparent from the following detailed description, the appended claims, and the accompanying drawings in which like reference numerals identify similar or identical elements.
The present invention is directed to three-stage switch fabrics that are based on different configurations of three different types of devices: an input device, a crossbar device, and an output device. A particular configuration may involve multiple instances of each of these three different types of devices. Logically, for a particular three-stage switch fabric configuration, the one or more input devices form the input stage (also referred to as the 1st stage) of switch fabric, the one or more crossbar devices form the switching stage (also referred to as the 2nd or middle stage), and the one or more output devices form the output stage (also referred to as the 3rd stage).
Each input device performs a user port expansion function to connect an associated set of one or more ingress line cards (also referred to as ingress traffic management (TM) ports) to all crossbar devices in the switching stage. Each crossbar device performs switching functions that connect any ingress TM port of any input device to any egress TM port of any output device. Each output device performs a user port contraction function to concentrate traffic from all crossbar devices and forward the traffic to an associated set of one or more egress line cards (i.e., egress TM ports). Those skilled in the art will understand that, in certain implementations, a particular line card may function as both an ingress TM port and an egress TM port.
Input, crossbar, and output devices can be configured together in different ways using different numbers of the different types of devices to achieve a wide range of switch fabric architectures having a wide range of switch capacities. For example, one type of scalable switch fabric architecture can be achieved by stacking together one or more instances of a basic switch fabric module, where each module comprises a single input device, a single crossbar device, and a single output device.
As shown in
Both input and output devices 204 and 208 have embedded memory as shared buffers. They aggregate 80 Gbps of user data and support user ports up to OC-768c rate. In particular, data is stored in memory that is organized as routing queues. The queuing structure ensures that there is no head-of-line blocking and provides traffic control such as buffer allocation and back-pressure. The switching stage is based on crossbar and has an arbitration logic that resolves port contentions.
For the scalable architecture of
In one implementation of the present invention, for every crossbar device 206 (up to 32) there can be up to two input devices 204 and two output devices 208 for a maximum (64×32×64) switch fabric configuration having a 5-Tbps switch capacity with 2:1 speed-up. Further upgrade for a maximum system capacity of 10 Tbps can be achieved by replacing the crossbar devices with high-capacity 256×256 crossbar elements. For this embodiment, the theoretical upper limit is 40 Tbps.
The fabric interface is preferably based on high-speed SERDES (serializer/de-serializer) technology, where each high-speed SERDES is embedded on chip. Both clock and synchronization are preferably embedded onto the data stream and recovered by integrated CDR (clock and data recovery) circuitry. The port rate is scalable from OC-12c/GbE rate to OC-768c rate. The high-speed serial signal I/O (input/output) can drive the backplane (or cables or optical transceivers directly) for inter-module and inter-shelf connectivity. This very high integration level and advanced system architecture provides a superb system solution for building very large scale switching or compact switching platforms. Further, the fabric interface supports multiple traffic classes and end-to-end flow control. Depending on the implementation, control information can be carried in-band or out-of-band or a combination of the two.
The out-of-band interface between input device 204 and its corresponding output device 208 includes the 1st stage back-pressure signal (1SBKP), the egress back-pressure signal (EBKP), the 3rd stage back-pressure signal (3SBKP), and the grant signal (GRANT).
When the cell buffer memory in the input device is full, the 1SBKP signal needs to be sent to ingress TM port. This is achieved by sending the 1SBKP signal out-of-band to the paired output device 208 (as shown in
When an egress TM port interface buffer is full, the EBKP signal is generated and passed to the ingress TM port. The EBKP signal is embedded in the header of an ingress cell. The input device extracts the EBKP signal and passes it out-of-band to the corresponding output device (as shown in
When the corresponding cell buffer memory in the output device is full, the 3SBKP signal is sent out-of-band to the input device (as shown in
The in-band GRANT signal from a crossbar device is extracted by the corresponding output device and passed out-of-band to the input device (as shown in
EMIF represents the external memory interface between input device 204 and a multicast look-up table (MCLKT) stored in external memory 502. MPI represents the microprocessor interface via which control and status information is exchanged between the input device and an external processor.
Referring again to
Cells from input ports are multiplexed at input mux 406 into a single cell stream and written to the cell buffer memory (BMEM) 408. The input mux also extracts routing information from the cell header and forwards it to the queue manager. Using this routing information, the queue manager supplies the write address of the cell to the BMEM.
The BMEM is a shared memory that is organized into a number of routing queues. The queue manager manages queues. For a unicast cell, the routing queue number and its traffic class determine to which queue the cell is admitted. For a multicast cell, it is always admitted to a dedicated multicast queue. The multicast tree structure is retrieved from the multicast look-up table (MCLKT) stored in external memory 502. The queue manager also performs the 1st stage back-pressure threshold test function. When a queue exceeds its buffer allocation, an appropriate 1st stage back-pressure signal is generated and placed onto the 1SBKP interface.
When a routing queue is back-logged, the scheduler 418 will allocate service to the queue. A service request (i.e., bid) is sent to pending logic 420. This request is then embedded into the header of a cell to be taken to the middle stage. (Note that the payload of this cell comprises user data corresponding to a previously granted bid.) The pending logic then waits to see if a bid is granted or rejected. If the request is granted, then a valid grant signal is expected from the GRANT interface. This grant signal is then fed to the pending logic which will initiate a real service for the corresponding bid.
A real service triggers a dequeue operation by the queue manager. One cell is read from the BMEM at a time. The BMEM read operations are performed by the queue manager and are scheduled so that every 40 cells from the BMEM will be de-multiplexed at output demux 410 to the 40 output port processors 412. The output demux converts a cell to the switch input port interface (SIPI) format. Switch requests and the 3rd stage back-pressure signals are inserted into the cell header. The assembled cells are then forwarded to the appropriate output port processor, which buffers them and forwards them to the output SERDES. Each output SERDES 414 converts cells to a serial bit stream and drives output buffers.
The presence of 40 SIPI outputs in each input device 204 and 40 SOPI inputs in each output device 208 enables each input and output device to be configured to up to 40 different crossbar devices 206. This extra capacity can be used for improving performance and/or implementation of certain redundancy or fault protection schemes, depending on system design requirements.
The interface between output device 208 and its corresponding input device 204 includes the 1st stage back-pressure signal (1SBKP), the egress back pressure signal (EBKP), the 3rd stage back-pressure signal (3SBKP), and the grant signal (GRANT), which were described previously in the context of
EMIF represents the external memory interface between output device 208 and the multicast look-up table (MCLKT) stored in external memory 702. MPI represents the microprocessor interface via which control and status information is exchanged between the output device and the external processor.
Referring again to
Cells from input ports are multiplexed at input mux 606 into a single cell stream and written to the cell buffer memory (BMEM) 608. The input mux also extracts routing information from the cell header and forwards it to the queue manager. Using this routing information, the queue manager supplies the write address of the cell to the BMEM.
The BMEM is a shared memory that is organized to a number of routing queues. The queue manager manages queues. For a unicast cell, the routing queue number and its traffic class determine to which queue the cell is admitted. For a multicast cell, it is always admitted to a dedicated multicast queue. The multicast tree structure is retrieved from the multicast look-up table (MCLKT) stored in external memory 702. The queue manager also performs the 3rd stage back-pressure threshold test function. When a queue exceeds its buffer allocation, an appropriate 3rd stage back-pressure signal is generated on the 3SBKP interface.
When a routing queue is back-logged, the scheduler 618 will allocate service to the queue. A service results in a dequeue operation. One cell is read from the BMEM at a time. The BMEM read operations are performed by the queue manager and are scheduled so that every 32 cells from the BMEM will be de-multiplexed at output demux 610 to the 32 output port processors 612. The output demux converts a cell to the fabric output port interface (FOPI) format. The 1st stage back-pressure signals are inserted into the cell headers. The assembled cells are then forwarded to the appropriate output port processor, which buffers them and forwards them to the output SERDES. Each output SERDES 614 converts cells to a serial bit stream and drives output buffers.
MPI represents the microprocessor interface via which control and status information is exchanged between the crossbar device and the external processor.
Referring again to
The crossbar block 808 is of output mux type. It has an array of output multiplexors (not shown) that select input ports, where the output multiplexors are configured on a per cell basis. The configuration data is stored in configuration register 818 and loaded to the crossbar device at the cell slot clock rate. Multiple configuration register files are stored to match the request/grant latency and input processor pipe-line operation.
The arbitration logic preferably performs a dual-priority arbitration function that resolves any collisions. All fabric ports are placed in a strict priority order. When a collision occurs, the highest priority port wins and its priority is degraded to the lowest. There are also two traffic priority levels. High-priority traffic always wins over low-priority traffic.
A crossbar device can be configured to different modes, e.g., from a single (64×64) matrix to 32 (2×2) matrices. A matrix configuration should match the system configuration. For example, in an 80-Gbps switch fabric configuration, one crossbar device may be used to connect two input devices and two output devices. In that case, the crossbar device should be configured as 32 (2×2) switches.
In preferred embodiments, a switch fabric of the present invention is a cell-switching engine handling fixed-size switching cells. A switching cell has a header and a payload. The payload size is programmable (e.g., either 64, 72, or 80 octets), while the header size is fixed (e.g., 12 octets). The term “programmable” implies that, in this case, the particular payload size is selected when the switch fabric is initially configured. Once the switch fabric is configured, the payload size remains fixed for all subsequent switch fabric operations. The selection of the payload size depends on the application. A smaller payload size is best suited for traffic with a large volume of small packets. ATM switching is a good example of when to choose a small payload size (e.g., 64 octets). A larger payload size as a better bandwidth utilization when large packets make up a large percentage of traffic. The switching cell header contains information used only by the switch fabric.
Protocol independence is achieved by encapsulating user data packets into the payload of switching cells. This encapsulation function is provided by network processors on the line cards. When user data packets are larger than the switching cell size, the encapsulation function involves dividing each user data packet into two or more different switching cells.
When converting variable-sized packets into fixed-size switching cells, the “N+1” problem is well known (i.e., where the last fixed-size switching cell for a particular user data packet has only a single unit (e.g., one byte) of user data). Assuming that a payload size of N data units is chosen for the switch fabric, any packet with the size ranging from N+1 to 2N requires two switching cells. In order to mitigate this problem and to provide additional bandwidth for the switching cell headers, the fabric interface bandwidth is expanded by 2-to-1 and every fabric port can provide twice the bandwidth of a user port.
As described previously in the context of the three-stage switch fabric 200 of
For example, in a (32×16×32) configuration having 32 input devices 204, 16 crossbar device 206, and 32 output devices 208, in which each input device had all 32 of its input ports configured to ingress line cards and each output device had all 32 of its output ports configured to egress line cards, there would be a total of 1024 configured output ports in the output stage (i.e., 32 output devices, each with 32 output ports configured to egress line cards). In that case, in order to provide a non-blocking switch fabric, each of the 1024 input routing queues in each input device would be logically associated with a different one of the 1024 output ports. In order to be able to reach any of the 1024 output ports, each of the input ports in each input device would be able to store user data to any of the 1024 input routing queues of that input device.
As another example, in a (32×16×32) configuration in which each input device had only 16 of its input ports configured to ingress line cards and each output device had only 16 of its output ports configured to egress line cards, there would be a total of 512 configured output ports in the output stage (i.e., 32 output devices, each with only 16 output ports configured to egress line cards). In that case, a different pair of the 1024 input routing queues in each input device could be logically associated with a different one of the 512 configured output ports, while still providing a non-blocking switch fabric. Since there could now be two different input routing queues for each configured output port, according to the present invention, the logical associations between the 16 configured input ports (e.g., “Input Port 1” to “Input Port 16”) and the 512 pairs of input routing queues of each input device can be programmed when the switch fabric is configured. For example, in a particular input device, for a particular pair of input routing queues (e.g., “Queue A” and “Queue B”) associated with a particular one of the 512 configured output ports (e.g., “Output Port I”), a first subset of the 16 configured input ports (e.g., Input Ports 1–10) could be programmed to store all of their user data destined for Output Port I into Queue A, while a second subset of the 16 configured input ports (e.g., Input Ports 11–16) could be programmed to store all of their user data destined for Output Port I into Queue B. Alternatively, the 16 configured input ports could be programmed to store one set of user data destined for Output Port I into Queue A and to store another set of user data destined for Output Port I into Queue B, thereby doubling the number of user traffic classes.
As yet another example, in a (4×2×4) configuration in which each input device had only 16 of its input ports configured to ingress line cards and each output device had only 16 of its output ports configured to egress line cards, there would be a total of 64 configured output ports in the output stage (i.e., 4 output devices, each with 16 output ports configured to egress line cards). In that case, a different set of 16 of the 1024 input routing queues in each input device could be logically associated with a different one of the 64 configured output ports, while still providing a non-blocking switch fabric. Since there could now be as many as 16 different input routing queues for each configured output port, according to the present invention, the logical associations between the 16 configured input ports (e.g., “Input Port 1” to “Input Port 16”) and the 64 sets of input routing queues of each input device can be programmed when the switch fabric is configured. For example, Input Port 1 could be programmed to write to Queues A and B (e.g., depending on the priority of its user data), Input Port 2 could be programmed to write to only Queue B, Input Port 3 could be programmed to write to Queues A and C, etc. Those skilled in the art will understand that there is great variability in how these different logical associations could be programmed, including different combinations involving individual input ports being able to store data into two or more different input routing queues and two or more input ports being able to store data into a single input routing queue.
In general, when a three-stage switch fabric of the present invention is initially configured, each input port in each input device can be programmed to be logically associated with one or more routing queues in that input device, where the one or more routing queues are selected from a group of routing queues whose number is dictated by the number of input ports of that input device that are configured to ingress line cards and the total number of configured output ports in the output stage of the switch fabric.
Similarly, when the switch fabric is initially configured, each output port in each output device can be programmed to be logically associated with one or more routing queues in that output device, where the one or more routing queues are selected from a group of routing queues whose number is dictated by the number of output ports of that output device that are configured to egress line cards and the total number of configured input ports in the input stage of the switch fabric. The programmability of the logical association between output ports and routing queues in output devices is analogous to that for the input devices.
For a given configuration, the mapping between routing queues in the input devices and routing queues in the output devices is determined based on the logical associations (a) between the input ports and the routing queues in the input devices and (b) between the output ports and the routing queues in the output devices.
The programming of one input port of an input device is independent of the programming of the other input ports of that input device. Similarly, the programming of the input ports of one input device in an input stage is independent of the programming of the input ports of the other input devices in the input stage. Analogous independence exists for the programming of the output ports of the output devices in the output stage.
This ability to program the logical associations between ports and queues in the input and output devices increases the flexibility provided by switch fabric architectures of the present invention. As described earlier, the programmability enables two or more inputs to be logically grouped together to form a high-rate logical input port. Similarly, the programmability enables two or more outputs at an output device to be logically grouped together to form a high-rate logical output port.
When an input port is associated with two or more different routing queues in an input device, different sets of user data received at that input port can be separated and stored in different routing queues. This allows the different sets of user data from the same source to be treated differently, for example, based on their different destinations and/or to provide different treatment for different quality-of-service (QoS) levels and/or to achieve isolation. For example, if different types of user data are stored in different routing queues and one of the routing queues is full, then isolation can be achieved by sending back-pressure information to the source to stop sending the corresponding type of user data, while allowing the other types of user data to continue to be transmitted from that source for storage in the other routing queues associated with that same input port.
Servicing of (i.e., data removal from) those multiple routing queues in both the input and output devices (as controlled by the schedulers in those devices) is preferably sequenced in order to preserve the order in which the user data arrived at the switch fabric.
Queue Structure, QoS Schedulers, and Cell Routing
The switch fabric of the present invention provides a unique queuing structure that supports advanced traffic management features such as end-to-end quality of service (QoS), traffic isolation, and fair use of the system bandwidth. The queuing structure supports a large number of routing queues within the fabric that can be organized, according to the system configuration, to provide simple VOQ (virtual output queue) implementation or input/output pair queuing (IOPQ). For scalability, a routing queue numbering scheme is used instead of a port bitmap. Further, there are two sub-queues per routing queue for scheduling purposes. Delay-sensitive traffic can be switched with highest priority to minimize the latency. Additionally, virtual time-division multiplexing (TDM) pipes can be set up between ports to provide guaranteed bandwidth and minimize the delay variation. Work conserving fair queuing scheduling allocates bandwidth for the best-effort traffic that is treated with low priority through the switch fabric.
Bid and Arbitration Mechanism
The bid and arbitration mechanism is responsible for requesting slots to transmit cells from the input devices to the output devices and resolving any collisions in the crossbar devices. In a configuration in which there is only one link from each crossbar device to each output device, a collision refers to a situation in which two or more input devices transmit bids (i.e., service requests) to the same crossbar device during the same time slot (also referred to as a cell slot) to request connections to the same output device. For configurations in which two or more links connect a particular crossbar device to a particular output device, a collision will occur when the number of bids received by that crossbar device for that output device during the same time slot is greater than the number of links between that crossbar device and that output device.
Input Device Randomizer
In the embodiment of
For example, if each input device is configured with 32 connection links to the switching stage (e.g., one per crossbar device in a (64×32×64) configuration), then, for every 32 bids, the PRN generator generates a different pseudo-random number (PRN) in the range from 1 to 32. The number PRN is then used to map the 32 different bids to the 32 different connection links. For example, in one implementation, the number PRN is the link number for Bid 1 (i.e., the first of the 32 bids), the number PRN+1 is the link number for Bid 2, etc., in a circular manner where link number 1 follows link number 32 for PRN values other than 1. Since there is latency between a bid and its corresponding grant signal, a grant for a bid is expected on the middle stage output link corresponding to the same middle stage input link on which the bid was sent. To match the bid/grant sequence, the input device remembers the sequence of PRNs generated by the PRN generator. When the input stage has multiple input devices, different PRN seeds should be loaded into the PRN generators of the different input devices to avoid undesirable synchronization between those PRN generators.
Bid Data Format
Bids are inserted into the headers of cells transmitted from the input device to the crossbar devices. (Note that, since bids are for future routing of cells, the payload of each cell carrying a bid corresponds to earlier transmitted and granted bid.) In the embodiment of
Both the requested 3rd stage device (RDESTD) and the requested 3rd stage port (RDESTP) are directly extracted from the requested destination field REQDEST of the cell header. REQDEST is decomposed to RDESTD and RDESTP using the configured format. RDESTD and RDESTP are checked against global parameters MAXDEVICE and MAXPORT, respectively. If a range violation is detected, the request is rejected. The 3rd stage routing queue number R3QN is extracted from the requested routing queue number field REQRQN of the cell header. RSRCD indicates the input device that originated the request. It is derived from the input port number and used for priority selection and for arbitration.
3rd Stage Back-pressure Test
In preferred implementations, the bid arbitrator in each crossbar device takes into account the 3rd stage back-pressure status. A congested output device is removed from the arbitration because a cell cannot be sent to the congested output device. Each crossbar device maintains a 3rd stage back-pressure status table. Before arbitration is performed, a valid bid is examined against the back-pressure status. There are separate back-pressure status tables for unicast and multicast transmissions.
The unicast back-pressure status is organized in three levels: per output device, per destination fabric port, and per queue. Every level of back-pressure is further divided into two priorities.
The 3rd stage back-pressure is selective. When a 3rd stage buffer queue is congested, only the traffic that is destined to the congested queue is back-pressured, while traffic that is destined to other queues and ports in the same output device may still go through the fabric.
The first level of the back-pressure is the 3rd stage device level back-pressure 3DBKP. This level controls the traffic to an output device. An output device is back-pressured if its buffer memory is full. The second level of the back-pressure is the 3rd stage port back-pressure 3PBKP. This level controls the traffic to an output device port. An output device port is back-pressured if the buffer memory allocation for the port is used up. The third level of the back-pressure is the 3rd stage routing queue level back-pressured 3QBKP. This level controls the traffic to an output device routing queue. When a queue fill level reaches its back-pressure level, traffic sources (e.g., 1st stage queues, ports, or devices) are stopped.
For a unicast request (BIDCH=0/1), the back-pressure test is as follows:
For multicast requests, an output device has a single multicast queue shared by all ports. There is a single level multicast back-pressure control on the per output device basis. The multicast back-pressure status table MCBKPSBT is organized as a one-dimensional table, as shown in
In one possible implementation, both the per queue and the per port back-pressure information are used in the input stage. In another possible implementation, only the per queue back-pressure information is used in the input stage, while the per port back-pressure information is used in the switching stage.
Crossbar Contention and Arbitration Algorithms
While the number of links between the input stage and the output stage gives the total bandwidth internal to the fabric devices, the actual user data throughput depends on the traffic distribution and the applied contention/arbitration algorithm.
Random Bid/Random Arbitration
The arbitration scheme of
Bids are first sorted according to their traffic class. They are then sorted according to the requested output devices. Within each crossbar device, the sorted bids are stored in bid registers that are organized per output device, as shown in
The results from the bid arbitration (i.e., the grant/rejection signals) are sent from each crossbar device to the appropriate input devices via the output devices that are paired with those input devices.
In the maximum 64×32×64 architecture (i.e., a switch fabric having 64 input devices, 32 crossbar devices, and 64 output devices), at each time slot, each input device sends up to 32 bids, no more than one to each of the 32 different crossbar devices, to contend for the cross-connects in those crossbar devices. These 32 requests are essentially randomly distributed over the 32 crossbar devices. At each time slot, each crossbar device arbitrates over its (up to 64) received bids, which includes at most one bid from each input device, to determine which bids to grant (accept) and which bids to deny (reject). The grant/deny decisions are encoded in the headers of cells transmitted back to the corresponding input devices. The bid arbitration processing performed by each crossbar device follows the rules that, for each requested output device, (1) if the destination port and/or the destination queue is back-pressured, then the bid is rejected, (2) contending bids with lower priority (e.g., GBW class) are rejected in favor of those with higher priority (e.g., EBS (excessive bandwidth scheduler) class), and (3) for each output device, one of the remaining bids (necessarily of the same priority) is randomly chosen (for fairness) to be granted. In effect, this random/random scheme approximates the random bids where the requests from the input devices are randomized from each other.
Note that, since a particular input device can send several cells belonging to one flow in one cell slot (via different crossbar devices), the issue of keeping the flow cell sequence throughout the switch fabric is addressed by allocating the cells from the same input device queue to crossbar devices (that accepted the bids from this queue) in prescribed order in each cell slot. Namely, if there is more than one cell from a given queue to be sent in a given cell slot, the first cell from the queue is allocated to the eligible crossbar device with the lowest index, the next cell is allocated to the eligible crossbar device with the second lowest index, etc. Then, in the output device, the cells arriving in a given cell slot from the crossbar devices are transferred into the cell memory following the crossbar index order as illustrated in
The worst case of the maximum throughput occurs when the crossbar device is subject to symmetric uniform traffic. In that case, the maximum throughput Umax of the crossbar device is given by Equation (1) as follows:
U max=1−(1−1/M)M (1)
where M is the number of input devices, where throughput corresponds to the fraction of bids that are accepted. For other distributions, capacity bounds can also be derived. Note that, for large M (e.g., 64), the worst-case maximum utilization is about 63.5%. With the 2-to-1 speed-up at the fabric interfaces, this may or may not be sufficient to match the line rate at the port card, depending on the application and packet size distribution. There are a number of ways of improving the data throughput. One way is to limit the aggregate traffic bandwidth per device or increase the number of links between the stages. Another approach is to improve upon the bid and arbitration scheme and the architecture to achieve a higher efficiency.
Persistent Bid/Random Arbitration
One alternative bid/arbitration scheme is based on a persistent bid/random arbitration approach.
Since the scheduler in each input device can handle as many as 32 scheduling decisions in one cell slot (i.e., the maximum number of links to crossbar devices in the switching stage), the congestion point is at the pending logic, and flow control is needed to suspend the scheduler properly. One possible flow control scheme is the window flow control with a constant window size. Other issues to consider are the disposition of the cell priority, the denied bids, and the choice and manner that the pending logic makes bids to the crossbar devices among the available cells.
In a particular implementation that supports different traffic classes, when a bid for a low-priority routing queue is rejected, a low-priority cell in that queue is designated for rescheduling for service by the scheduler. In order to preserve the proper cell sequence, the cell that is to be rescheduled is the last low-priority cell in the queue. For example, if a routing queue has two low-priority cells and a bid for that queue is rejected, then the second low-priority cell is to be rescheduled, while another bid is transmitted for the first low-priority cell. If that second bid is also rejected, then the first low-priority cell will also be rescheduled for service. When the two low-priority cells are again scheduled for service, the order of the cells is preserved for proper sequencing.
Arbitration with Buffered Crossbar Architecture for Multiple Grants
Assume that there are L bids for a particular destination in slot n. One possible algorithm is as follows:
Note that, in the buffered crossbar architecture of
The output device design is modified by introducing, for each input device, an additional buffer, working as a shift register, at the output device input link ports to provide resequencing of cells routed by different crossbar devices. The additional buffer should have (K+1)-cell capacity (the same number as in the crossbar device). This solution is illustrated in
In the worst case, the output stage has to transfer up to 32*(K+1) cells from the output device input port buffers to the memory in one cell slot. If this is not possible due to physical limitations, an extended version can be used where the output device input port buffers are extended to negative buffer cell indexes. In this case, all cells with delay equal to or larger than the corresponding buffer cell indexes are eligible for transfer. Then, the algorithm works the same way as before with the exception that the cells with the largest difference between the delay index and the buffer cell index are transferred first (across all buffers). A buffer capacity of 2*(K+1) provides that all cells will be transferred without any loss if the transfer speed is the same as in the design of
In certain implementations that support different traffic classes, a parameter P can be set to reserve a specified number of spaces in each crossbar buffer for high-priority cells. For example, assume a situation in which (a) each crossbar buffer can hold five cells, (b) the parameter P is set to three, and (c) a particular buffer currently has two low-priority cells and one high-priority cell. In that case, the two remaining spaces in the buffer are reserved for high-priority cells, and the buffer will be able to accept additional low-priority cells only after one or both of the currently stored low-priority cells are transmitted to the corresponding output device.
Arbitration with Buffered Crossbar Architecture for Reconsideration of Non-granted Bids
Another type of bid arbitration involves the ability to reconsider, during the next time slot, one or more bids that are not granted during the current time slot. To implement such an arbitration scheme, each crossbar device 206 is implemented with a number of (e.g., 64) bid buffers, where each bid buffer is able to store a specified number of bids for reconsideration during the next time slot. Depending on the implementation, a given bid can be reconsidered a specified number of times before determining that the bid should be rejected. When arbitrating a collision during the subsequent time slot, preference can be given to bids that are being reconsidered as opposed to newly received bids that are being considered for the first time.
In particular, collision resolution is performed to grant at most one bid, taking into account any new bids that just arrived for the current time slot as well as any old bids from the previous time slot that were stored in the corresponding bid buffer (step 2402 of
Depending on the implementation, a bid that is not granted during the current time slot may be reconsidered a specified number of times. For example, in one implementation, a bid may be reconsidered only once. In that case, if a bid is not granted during either the time slot at which it arrived at the crossbar device or the next time slot, then such a bid will be rejected as being too old during that second time slot (step 2404). In other implementations, bids may be reconsidered a maximum number of times more than one, including an unlimited number of times.
Any remaining non-granted (i.e., which are not too old) are then stored in the corresponding bid buffer for reconsideration during the next time slot (step 2406). Depending on the particular implementation, the size of the bid buffers may be limited such that there may not be enough room in a bid buffer to store all of the non-granted bids that are not too old. In that case, one or more non-granted bids that are not too old may need to be rejected in step 2406. Depending on the implementation, the age of such bids may be taken into account to determine which bids to store and which bids to reject. For example, newer bids may be given priority over older bids, or vice versa. Alternatively or in addition, random selection can be implemented to determine which bids to store and which bids to reject.
In a switch fabric that allows non-granted bids to be reconsidered during the next time slot, the pending logic in each input device (which is responsible for generating bids) should wait until it receives either a grant or a rejection signal for a previously transmitted service request for a particular routing buffer before deciding how to proceed for that routing buffer.
Permutations of Various Scheme Features
The previous sections describe different features involved in different types of bid/arbitration schemes. Some of these were random bidding, persistent bidding, random arbitration, arbitration with multiple grants, and arbitration with reconsideration of non-granted bids. For different implementations of the present invention, two or more of these different features can be combined in a single implementation. In addition to the random bidding/random arbitration and persistent bidding/random arbitration combinations described previously, other permutations are also possible. For example, in combination with either random or persistent bidding, the arbitration scheme can be based on either arbitration with multiple grants or arbitration with reconsideration of non-granted bids.
In addition, either random or persistent bidding can be combined with an arbitration scheme that allows both multiple grants and reconsideration of non-granted bids. In such a scheme, when collision occurs, two or more grants are allowed for each time slot. If more bids are involved in a collision than the maximum number of grants allowed (or, in particular, the number of available positions in the crossbar device's cell buffer), then one or more additional bids that cannot be granted during the current time slot, may be buffered in the crossbar device's bid buffer for reconsideration during the next time slot (along with any newly arrived bids).
For example, assume a situation in which (1) each cell buffer can hold up to six grants, (2) a particular cell buffer already has two grants from a previous time slot, and (3) the current time slot has a seven-way collision. In that case, since there is room in the cell buffer for another four grants, four of the seven contending bids can be granted and buffered in the cell buffer during the current time slot. Depending on their age, one or more of the remaining three contending bids (i.e., the non-granted bids) may be stored in the bid buffer for reconsideration during the next time slot. Any bids that are neither granted nor stored for reconsideration are determined to be rejected, and corresponding rejection signals are sent back to the appropriate input devices for those bids.
Due to the additional buffering introduced at the crossbar devices, the bid rejections caused by contention for the same output device are dramatically reduced and consequently a significant improvement in maximum utilization of the crossbars is achieved.
As shown in Table I, almost full crossbar capacity (99%) is achieved for K=30 cells. This corresponding to 55.9% gain over the no-buffer case. Nevertheless, to reduce both the complexity of the implementation and the additional delay introduced by buffering, the maximum number of grants per slot can be limited and smaller buffers can be used. For example, the case of K=5 with a maximum of 3 grants per time slot gives 91.6% throughput, corresponding to a 44.3% gain. If the maximum number of grants per time slot is further reduced to 2, then the throughput drops to 88.1%, corresponding to a 38.8% gain. Still, these numbers may be sufficient depending on what is considered the worst traffic pattern.
To provide a data view that would facilitate the choice of maximum number of grants and the buffer capacity,
Back-pressure and Flow Control
Back-pressure is used for end-to-end flow control among switching elements and the traffic management line cards to maximize the usage of buffer memory. Preferred embodiments of the present invention rely on an efficient back-pressure (BP) and fast flow control scheme to achieve a highly scalable, lossless switch fabric architecture. Further, to support QoS classes, embodiments of the present invention regulate and control the allocation of the bandwidth across the switch fabric to different flows. Back-pressure is preferably applied selectively only to the traffic destined to congested ports. Furthermore, the switch fabric preferably allows selective back-pressure to be asserted on any traffic bound for an output device that has reached its buffer allocation limit.
Preferred back-pressure and flow control schemes involve BP mechanisms (a) from the egress line cards to the output devices, (b) from the output devices to the input devices, (c) from the input devices to the ingress line cards, and (d) from the output devices to the crossbar devices.
There are three BP mechanisms from the output devices to the input devices: per queue, per port, and per device. The objective of the per queue BP mechanism is to block a routing queue in an input device when its corresponding queue in an output device exceeds a first specified threshold. The objective of the per port BP mechanism is to block all the input device queues that are associated with output device queues associated with a single port for which the total number of cells in the output device queues exceeds a second specified threshold. Finally, the objective of the per device BP mechanism is to block all the input device queues that are associated with queues in an output device for which the total number of cells in all of the queues in that output device exceeds a third specified threshold.
The per port BP mechanism is an indirect BP mechanism, since the per port BP states are not sent from the output devices directly to the input devices. Rather, the BP state for each port is delivered to all of the crossbar devices. The bid arbitrator in each crossbar device uses these per port states in determining whether to accept or reject its current set of bids received from the input devices, where a bid is rejected if the corresponding BP state is active, indicating that the port is not available to receive any additional cells at this time. The per port BP states are sent periodically using dedicated in-band signaling channels on the following path: output device=>out-of-band=>paired input device=>in-band =>all crossbar devices. In particular, the per port BP states are transmitted out-of-band from the output device to its paired input device (e.g., using 3SBKP in
One possible approach would be to apply the same solution for both the per port and the per queue BP mechanisms. There are two drawbacks to such an approach. First, in a well-configured system, the per queue BP will be activated significantly more often than the per port BP. This feature indicates that some fabric bandwidth can be underutilized due to the fact that bids are rejected due to the active per queue BP. This loss could be as high as 50% under some specific traffic scenarios. Second, since there are typically many more queues than ports, delivery of per queue BP states would take much more time than delivery of the per port BP states.
In order to avoid these problems, in preferred embodiments, the per queue BP states are delivered directly to the corresponding input devices using dedicated in-band signaling channels in the headers of cells transmitted between input and output devices via crossbar devices. The signaling path is as follows: output device=>out-of-band=>paired input device=>in-band=>corresponding output device via crossbar device=>out-of-band=>paired input device. In particular, the per queue BP states are transmitted out-of-band from the output device to its paired input device (e.g., using 3SBKP in
This architecture has certain advantages. Among them is that, due to the low per queue BP, per queue QoS features can be supported effectively in the fabric. In addition, since the BP state is delivered directly to the corresponding input device, there is no loss of internal fabric bandwidth since back-pressured flows would not contend for the bandwidth.
In terms of the reliability consideration of the per queue BP signaling, there are at least two solutions. One solution is to use redundant channels by introducing two channels per input device/output device pair. This solution increases the per queue BP latency from 2 to 4 slots. Another solution assumes that, since, from the reliability viewpoint, the most important thing is to provide the lossless feature, one can relegate this feature to the per port and per device BP mechanisms. In this case, once there is a link failure, the BP states in the corresponding input device are set to zero (i.e., no back-pressure).
With judicious setting of the back-pressure thresholds per device, per port, and per queue, an efficient lossless fabric operation can be achieved with per queue QoS and minimum buffers on device.
Although the present invention has been described in the context of electronic switch fabrics, those skilled in the art will understand that the present invention can also be implemented in the context of optical switch fabrics as well as switch fabrics that involve both electronic and optical processes.
The present invention may be implemented as circuit-based processes, including possible implementation on and fabrication as one or more integrated circuits. As would be apparent to one skilled in the art, various functions of circuit elements may also be implemented as processing steps in a software program. Such software may be employed in, for example, a digital signal processor, micro-controller, or general-purpose computer.
Unless explicitly stated otherwise, each numerical value and range should be interpreted as being approximate as if the word “about” or “approximately” preceded the value of the value or range.
It will be further understood that various changes in the details, materials, and arrangements of the parts which have been described and illustrated in order to explain the nature of this invention may be made by those skilled in the art without departing from the scope of the invention as expressed in the following claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5124978||Jan 3, 1991||Jun 23, 1992||Bell Communications Research, Inc.||Grouping network based non-buffer statistical multiplexor|
|US5157654 *||Dec 18, 1990||Oct 20, 1992||Bell Communications Research, Inc.||Technique for resolving output port contention in a high speed packet switch|
|US5255265||May 5, 1992||Oct 19, 1993||At&T Bell Laboratories||Controller for input-queued packet switch|
|US5361255||Apr 29, 1991||Nov 1, 1994||Dsc Communications Corporation||Method and apparatus for a high speed asynchronous transfer mode switch|
|US5440553||Dec 19, 1994||Aug 8, 1995||Widjaja; Indra||Output buffered packet switch with a flexible buffer management scheme|
|US5497369||Oct 31, 1991||Mar 5, 1996||Hewlett-Packard Company||Multicast switch circuits|
|US5500858 *||Dec 20, 1994||Mar 19, 1996||The Regents Of The University Of California||Method and apparatus for scheduling cells in an input-queued switch|
|US5537403||Dec 30, 1994||Jul 16, 1996||At&T Corp.||Terabit per second packet switch having distributed out-of-band control of circuit and packet switching communications|
|US5550815||Dec 30, 1994||Aug 27, 1996||Lucent Technologies Inc.||Apparatus and method for reducing data losses in a growable packet switch|
|US5555243||Dec 9, 1994||Sep 10, 1996||Fujitsu Limited||Self routing exchange and exchange system|
|US5612953||Oct 18, 1995||Mar 18, 1997||International Business Machines Corporation||Multi-media serial line switching adapter for parallel networks and heterogeneous and homologous computer systems|
|US5631902||Apr 28, 1994||May 20, 1997||Nec Corporation||Digital cross-connect system|
|US5689500||Jan 16, 1996||Nov 18, 1997||Lucent Technologies, Inc.||Multistage network having multicast routing congestion feedback|
|US5689505||Jan 16, 1996||Nov 18, 1997||Lucent Technologies Inc.||Buffering of multicast cells in switching networks|
|US5754120||Dec 21, 1995||May 19, 1998||Lucent Technologies||Network congestion measurement method and apparatus|
|US5852601||Jun 9, 1997||Dec 22, 1998||Network Equipment Technologies, Inc.||Method and apparatus for reactive congestion control in an asynchronous transfer mode (ATM) network|
|US5938749||Jun 3, 1996||Aug 17, 1999||Whittaker Communications Inc.||Queue measurement apparatus and methodology|
|US6125112 *||Mar 23, 1998||Sep 26, 2000||3Com Corporation||Non-buffered, non-blocking multistage ATM switch|
|US6160812 *||May 4, 1998||Dec 12, 2000||Cabletron Systems, Inc.||Method and apparatus for supplying requests to a scheduler in an input buffered multiport switch|
|US6229812||Jun 11, 1997||May 8, 2001||Paxonet Communications, Inc.||Scheduling techniques for data cells in a data switch|
|US6243382||Oct 27, 1997||Jun 5, 2001||Fujitsu Limited||Interfacing to SAR devices in ATM switching apparatus|
|US6324165||Sep 5, 1997||Nov 27, 2001||Nec Usa, Inc.||Large capacity, multiclass core ATM switch architecture|
|US6430181||Jan 15, 1999||Aug 6, 2002||Lsi Logic Corporation||Crossbar switch fabric with deterministic maximal scheduling of connection requests with strict scheduling prioritization derived from requested service delays|
|US6438134||Aug 24, 1998||Aug 20, 2002||Alcatel Canada Inc.||Two-component bandwidth scheduler having application in multi-class digital communications systems|
|US6519225 *||May 14, 1999||Feb 11, 2003||Nortel Networks Limited||Backpressure mechanism for a network device|
|US6560230||May 21, 1999||May 6, 2003||Redback Networks Inc.||Packet scheduling methods and apparatus|
|US6570873 *||Nov 12, 1999||May 27, 2003||Nec Corporation||System and method for scheduling reservation of traffic with priority|
|US6762995||Jun 14, 2000||Jul 13, 2004||3Com Corporation||Network switch including hysteresis in signalling fullness of transmit queues|
|US6778536||Mar 28, 2000||Aug 17, 2004||Synchrodyne Networks, Inc.||Combined wavelength division multiplexing, time division multiplexing, and asynchronous packet switching with common time reference|
|US6829237||Oct 9, 2001||Dec 7, 2004||Irvine Sensors Corporation||High speed multi-stage switching network formed from stacked switching layers|
|US6865154||Feb 25, 1999||Mar 8, 2005||Enterasys Networks, Inc.||Method and apparatus for providing bandwidth and delay guarantees in combined input-output buffered crossbar switches that implement work-conserving arbitration algorithms|
|US6868084||Sep 27, 2001||Mar 15, 2005||Teak Networks, Inc||Strictly nonblocking multicast multi-stage networks|
|US6870831||May 1, 2001||Mar 22, 2005||Pmc-Sierra, Inc.||Flexible, self-aligning time and space switch fabrics|
|US6907041 *||Mar 7, 2000||Jun 14, 2005||Cisco Technology, Inc.||Communications interconnection network with distributed resequencing|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7212523 *||Mar 27, 2003||May 1, 2007||Applied Micro Circuits Corporation||Pipeline architecture for the design of a single-stage cross-connect system|
|US7292594 *||Jan 9, 2003||Nov 6, 2007||Lsi Corporation||Weighted fair share scheduler for large input-buffered high-speed cross-point packet/cell switches|
|US7304988 *||Mar 28, 2003||Dec 4, 2007||Applied Micro Circuits Corporation||Technique for building a large single-stage cross-connect using multiple devices without interleaving|
|US7324541 *||Dec 22, 2003||Jan 29, 2008||Intel Corporation||Switching device utilizing internal priority assignments|
|US7359396 *||Oct 9, 2003||Apr 15, 2008||Nortel Networks Limited||Optical-core network with selective signal queueing|
|US7408947 *||Jan 6, 2005||Aug 5, 2008||Enigma Semiconductor||Method and apparatus for scheduling packets and/or cells|
|US7542464 *||May 14, 2004||Jun 2, 2009||Fraser Alexander G||Switching network|
|US7570654||Dec 22, 2003||Aug 4, 2009||Intel Corporation||Switching device utilizing requests indicating cumulative amount of data|
|US7623524||Dec 22, 2003||Nov 24, 2009||Intel Corporation||Scheduling system utilizing pointer perturbation mechanism to improve efficiency|
|US7675909 *||Dec 15, 2004||Mar 9, 2010||Tellabs Operations, Inc.||Method and apparatus for horizontally slicing a multi-stage switch fabric|
|US7688815 *||Aug 17, 2004||Mar 30, 2010||BarracudaNetworks Inc||Method and system for a multi-stage interconnect switch|
|US7738472||Jul 24, 2008||Jun 15, 2010||Forestay Research, Llc||Method and apparatus for scheduling packets and/or cells|
|US7743186||Apr 27, 2007||Jun 22, 2010||Atmel Corporation||Serialization of data for communication with different-protocol slave in multi-chip bus implementation|
|US7761632||Apr 27, 2007||Jul 20, 2010||Atmel Corporation||Serialization of data for communication with slave in multi-chip bus implementation|
|US7769933||Apr 27, 2007||Aug 3, 2010||Atmel Corporation||Serialization of data for communication with master in multi-chip bus implementation|
|US7813348||Nov 3, 2004||Oct 12, 2010||Extreme Networks, Inc.||Methods, systems, and computer program products for killing prioritized packets using time-to-live values to prevent head-of-line blocking|
|US7814250 *||Apr 27, 2007||Oct 12, 2010||Atmel Corporation||Serialization of data for multi-chip bus implementation|
|US7830903||May 20, 2010||Nov 9, 2010||Forestay Research, Llc||Method and apparatus for scheduling packets and/or cells|
|US7843905 *||May 3, 2007||Nov 30, 2010||Nortel Networks Limited||Modular high-capacity switch|
|US8036128 *||Sep 28, 2007||Oct 11, 2011||Alcatel Lucent||Method for communicating backpressure messages in a data communications system|
|US8037060 *||Sep 29, 2006||Oct 11, 2011||A9.Com, Inc.||System and method for displaying columns of search results|
|US8072887 *||Feb 7, 2005||Dec 6, 2011||Extreme Networks, Inc.||Methods, systems, and computer program products for controlling enqueuing of packets in an aggregated queue including a plurality of virtual queues using backpressure messages from downstream queues|
|US8085658 *||Aug 20, 2010||Dec 27, 2011||Marvell International Ltd.||System and method of cross-chip flow control|
|US8284668 *||Feb 22, 2010||Oct 9, 2012||Fujitsu Limited||Packet transmission apparatus, line interface unit, and control method for packet transmission apparatus|
|US8284771 *||May 6, 2011||Oct 9, 2012||Telefonaktiebolaget L M Ericsson (Publ)||Run-time scalable switch fabric|
|US8346988||May 25, 2010||Jan 1, 2013||International Business Machines Corporation||Techniques for dynamically sharing a fabric to facilitate off-chip communication for multiple on-chip units|
|US8549207 *||Nov 18, 2010||Oct 1, 2013||The Regents Of The University Of Michigan||Crossbar circuitry for applying an adaptive priority scheme and method of operation of such crossbar circuitry|
|US8792516||Nov 1, 2010||Jul 29, 2014||Rockstar Consortium Us Lp||Modular high-capacity switch|
|US8868817||Apr 4, 2012||Oct 21, 2014||The Regents Of The University Of Michigan||Crossbar circuitry for applying an adaptive priority scheme and method of operation of such crossbar circuitry|
|US9514074||Jul 12, 2013||Dec 6, 2016||The Regents Of The University Of Michigan||Single cycle arbitration within an interconnect|
|US20030227932 *||Jan 9, 2003||Dec 11, 2003||Velio Communications, Inc.||Weighted fair share scheduler for large input-buffered high-speed cross-point packet/cell switches|
|US20040190503 *||Mar 27, 2003||Sep 30, 2004||Bansal Narendra K.||Pipeline architecture for the design of a single-stage cross-connect system|
|US20040190504 *||Mar 28, 2003||Sep 30, 2004||Bansal Narendra K.||Technique for building a large single-stage cross-connect using multiple devices without interleaving|
|US20040240437 *||May 14, 2004||Dec 2, 2004||Fraser Alexander G.||Switching network|
|US20050041637 *||Aug 17, 2004||Feb 24, 2005||Jan Bialkowski||Method and system for a multi-stage interconnect switch|
|US20050135355 *||Dec 22, 2003||Jun 23, 2005||Raman Muthukrishnan||Switching device utilizing internal priority assignments|
|US20050135356 *||Dec 22, 2003||Jun 23, 2005||Raman Muthukrishnan||Switching device utilizing requests indicating cumulative amount of data|
|US20050135398 *||Dec 22, 2003||Jun 23, 2005||Raman Muthukrishnan||Scheduling system utilizing pointer perturbation mechanism to improve efficiency|
|US20050207436 *||Mar 18, 2004||Sep 22, 2005||Anujan Varma||Switching device based on aggregation of packets|
|US20060126610 *||Dec 15, 2004||Jun 15, 2006||Tellabs, Inc.||Method and apparatus for horizontally slicing a multi-stage switch fabric|
|US20060146706 *||Jan 6, 2005||Jul 6, 2006||Enigma Semiconductor||Method and apparatus for scheduling packets and/or cells|
|US20070086464 *||Dec 8, 2006||Apr 19, 2007||Hosagrahar Somashekhar||Multi-service data transport architecture|
|US20070121660 *||Nov 6, 2006||May 31, 2007||Robert Ayrapetian||Load balanced architecture of cascading of MxM Ethernet packet based switches that supports up to 4 levels of QoS|
|US20070206946 *||May 3, 2007||Sep 6, 2007||Beshai Maged E||Modular High-Capacity Switch|
|US20080040232 *||Aug 14, 2006||Feb 14, 2008||Dirk Perchthaler||Method for performing a network auction|
|US20080084797 *||Oct 5, 2007||Apr 10, 2008||Kousei Sano||Optical head device and optical information device|
|US20080159145 *||Dec 29, 2006||Jul 3, 2008||Raman Muthukrishnan||Weighted bandwidth switching device|
|US20080270650 *||Apr 27, 2007||Oct 30, 2008||Atmel Corporation||Serialization of data for multi-chip bus implementation|
|US20080270656 *||Apr 27, 2007||Oct 30, 2008||Atmel Corporation||Serialization of data for communication with different-protocol slave in multi-chip bus implementation|
|US20080270667 *||Apr 27, 2007||Oct 30, 2008||Atmel Corporation||Serialization of data for communication with master in multi-chip bus implementation|
|US20090059942 *||Jul 24, 2008||Mar 5, 2009||Enigma Semiconductor, Inc.||Method and apparatus for scheduling packets and/or cells|
|US20090086636 *||Sep 28, 2007||Apr 2, 2009||Alcatel Lucent||Method for communicating backpressure messages in a data communications system|
|US20090262744 *||Apr 7, 2009||Oct 22, 2009||Fraser Alexander G||Switching network|
|US20100232449 *||May 20, 2010||Sep 16, 2010||Nielsen Jacob V||Method and Apparatus For Scheduling Packets and/or Cells|
|US20100238941 *||Feb 22, 2010||Sep 23, 2010||Fujitsu Limited||Packet transmission apparatus, line interface unit, and control method for packet transmission apparatus|
|US20110052191 *||Nov 1, 2010||Mar 3, 2011||Beshai Maged E||Modular High-Capacity Switch|
|US20110138098 *||Nov 18, 2010||Jun 9, 2011||The Regents Of The University Of Michigan||Crossbar circuitry for applying an adaptive priority scheme and method of operation of such crossbar circuitry|
|US20130311681 *||Jul 24, 2013||Nov 21, 2013||Gvbb Holdings S.A.R.L.||Changeable functionality in a broadcast router|
|US20150278135 *||Mar 26, 2014||Oct 1, 2015||International Business Machines Corporation||Oldest link first arbitration between links grouped as single arbitration elements|
|US20150278136 *||Apr 14, 2014||Oct 1, 2015||International Business Machines Corporation||Oldest link first arbitration between links grouped as single arbitration elements|
|U.S. Classification||370/231, 370/414, 370/388, 370/356, 370/360|
|International Classification||H04J1/16, H04Q11/04, H04L12/66, H04L12/56, H04Q11/00, H04L12/28|
|Cooperative Classification||H04Q11/0478, H04Q2011/0024, H04L2012/5635, H04L2012/5634, H04L49/1576, H04Q11/0005, H04L49/3081, H04Q2011/0052, H04L2012/5679|
|European Classification||H04L49/15E5, H04L49/30J, H04Q11/00P2, H04Q11/04S2|
|Feb 7, 2002||AS||Assignment|
Owner name: AGERE SYSTEMS GUARDIAN CORP., FLORIDA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:DELL, MARTIN S.;DZIONG, ZBIGNIEW M.;LI, WEI;AND OTHERS;REEL/FRAME:012556/0527;SIGNING DATES FROM 20020103 TO 20020115
|Jun 30, 2010||FPAY||Fee payment|
Year of fee payment: 4
|Jun 11, 2014||FPAY||Fee payment|
Year of fee payment: 8
|Apr 3, 2015||AS||Assignment|
Owner name: AVAGO TECHNOLOGIES GENERAL IP (SINGAPORE) PTE. LTD
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:AGERE SYSTEMS LLC;REEL/FRAME:035365/0634
Effective date: 20140804
|Feb 11, 2016||AS||Assignment|
Owner name: BANK OF AMERICA, N.A., AS COLLATERAL AGENT, NORTH
Free format text: PATENT SECURITY AGREEMENT;ASSIGNOR:AVAGO TECHNOLOGIES GENERAL IP (SINGAPORE) PTE. LTD.;REEL/FRAME:037808/0001
Effective date: 20160201